TF_TString* data = static_cast<TF_TString*>(TF_TensorData(t));
  for (int i = 0; i < batch_size; ++i) {
    TF_TString_Init(&data[i]);
    // The following input string length is large enough to make sure that
    // copy to tstring in large mode.
    std::string source =
        "This is the " + std::to_string(i + 1) + "th. data element\n";
    TF_TString_Copy(&data[i], source.c_str(), source.length());
  }

  Run({{func_feed, Int32Tensor(2)}}, func_op, 2 /*+=*/ + 5 + 1);

  // Verify input, output, and subset of edges in fdef.
  // The subset of edges we verify is a chain between feed1 and output to
  // make sure that the correct output is picked.
  tensorflow::FunctionDef fdef;
  ASSERT_TRUE(GetFunctionDef(func_, &fdef));
  VerifyFDefInputs(fdef, M({{"feed1"}, {"feed2"}}));
  VerifyFDefOutputs(fdef, M({{"test_loop_exit"}}));

# XLA uses different paths for platforms and crosstool_top.
build:cross_compile_linux_arm64_xla --config=cross_compile_base_xla
build:cross_compile_linux_arm64_xla --platforms=//tools/toolchains/cross_compile/config:linux_aarch64
build:cross_compile_linux_arm64_xla --crosstool_top=//tools/toolchains/cross_compile/cc:cross_compile_toolchain_suite

# RBE cross-compile configs for Linux Aarch64

  Returns:
    string of native cuda compute capabilities, separated by comma.
  """
  device_query_bin = os.path.join(
      environ_cp.get('CUDA_TOOLKIT_PATH'), 'extras/demo_suite/deviceQuery')
  if os.path.isfile(device_query_bin) and os.access(device_query_bin, os.X_OK):
    try:
      output = run_shell(device_query_bin).split('\n')
      pattern = re.compile('[0-9]*\\.[0-9]*')